For a while, some DeepMind employees referred to it as “Watermelon.” Later, executives called it “Mario.” Both code names meant the same thing: a secret plan to break away from parent company Google.
DeepMind feared Google might one day misuse its technology, and executives worked to distance the artificial-intelligence firm from its owner for years, said nine current and former employees who were directly familiar with the plans.
This included plans to pursue an independent legal status that would distance the group’s work from Google, said the people, who asked not to be identified discussing private matters.
One core tension at DeepMind was that it sold the business to people it didn’t trust, said one former employee. “Everything that happened since that point has been about them questioning that decision,” the person added.
Efforts to separate DeepMind from Google ended in April without a deal, The Wall Street Journal reported. The yearslong negotiations, along with recent shake-ups within Google’s AI division, raise questions over whether the search giant can maintain control over a technology so crucial to its future.
“DeepMind’s close partnership with Google and Alphabet since the acquisition has been extraordinarily successful — with their support, we’ve delivered research breakthroughs that transformed the AI field and are now unlocking some of the biggest questions in science,” a DeepMind spokesperson said in a statement. “Over the years, of course we’ve discussed and explored different structures within the Alphabet group to find the optimal way to support our long-term research mission. We could not be prouder to be delivering on this incredible mission, while continuing to have both operational autonomy and Alphabet’s full support.”
A is for Alphabetization
When Google acquired DeepMind in 2014, the deal was seen as a win-win. Google got a leading AI research organization, and DeepMind, in London, won financial backing for its quest to build AI that can learn different tasks the way humans do, known as artificial general intelligence.
But tensions soon emerged. Some employees described a cultural conflict between researchers who saw themselves firstly as academics and the sometimes bloated bureaucracy of Google’s colossal business. Others said staff were immediately apprehensive about putting DeepMind’s work under the control of a tech giant. For a while, some employees were encouraged to communicate using encrypted messaging apps over the fear of Google spying on their work.
At one point, DeepMind’s executives discovered that work published by Google’s internal AI research group resembled some of DeepMind’s codebase without citation, one person familiar with the situation said. “That pissed off Demis,” the person added, referring to Demis Hassabis, DeepMind’s CEO. “That was one reason DeepMind started to get more protective of their code.”
After Google restructured as Alphabet in 2015 to give riskier projects more freedom, DeepMind’s leadership started to pursue a new status as a separate division under Alphabet, with its own profit and loss statement, The Information reported.
DeepMind already enjoyed a high level of operational independence inside Alphabet, but the group wanted legal autonomy too. And it worried about the misuse of its technology, particularly if DeepMind were to ever achieve AGI.
Internally, people started referring to the plan to gain more autonomy as “Watermelon,” two former employees said. The project was later formally named “Mario” among DeepMind’s leadership, these people said.
“Their perspective is that their technology would be too powerful to be held by a private company, so it needs to be housed in some other legal entity detached from shareholder interest,” one former employee who was close to the Alphabet negotiations said. “They framed it as ‘this is better for society.'”
In 2017, at a company retreat at the Macdonald Aviemore Resort in Scotland, DeepMind’s leadership disclosed to employees its plan to separate from Google, two people who were present said.
At the time, leadership said internally that the company planned to become a “global interest company,” three people familiar with the matter said. The title, not an official legal status, was meant to reflect the worldwide ramifications DeepMind believed its technology would have.
Later, in negotiations with Google, DeepMind pursued a status as a company limited by guarantee, a corporate structure without shareholders that is sometimes used by nonprofits. The agreement was that Alphabet would continue to bankroll the firm and would get an exclusive license to its technology, two people involved in the discussions said. There was a condition: Alphabet could not cross certain ethical redlines, such as using DeepMind technology for military weapons or surveillance.
In 2019, DeepMind registered a new company called DeepMind Labs Limited, as well as a new holding company, filings with the UK’s Companies House showed. This was done in anticipation of a separation from Google, two former employees involved in those registrations said.
Negotiations with Google went through peaks and valleys over the years but gained new momentum in 2020, one person said. A senior team inside DeepMind started to hold meetings with outside lawyers and Google to hash out details of what this theoretical new formation might mean for the two companies’ relationship, including specifics such as whether they would share a codebase, internal performance metrics, and software expenses, two people said.
DeepMind and Google leadership at odds over AI ethics
From the start, DeepMind was thinking about potential ethical dilemmas from its deal with Google. Before the 2014 acquisition closed, both companies signed an “Ethics and Safety Review Agreement” that would prevent Google from taking control of DeepMind’s technology, The Economist reported in 2019. Part of the agreement included the creation of an ethics board that would supervise the research.
Despite years of internal discussions about who should sit on this board, and vague promises to the press, this group “never existed, never convened, and never solved any ethics issues,” one former employee close to those discussions said. A DeepMind spokesperson declined to comment.
DeepMind did pursue a different idea: an independent review board to convene if it were to separate from Google, three people familiar with the plans said. The board would be made up of Google and DeepMind executives, as well as third parties. Former US president Barack Obama was someone DeepMind wanted to approach for this board, said one person who saw a shortlist of candidates.
DeepMind also created an ethical charter that included bans on using its technology for military weapons or surveillance, as well as a rule that its technology should be used in ways that benefit society. In 2017, DeepMind started a unit focused on AI ethics research composed of employees and external research fellows. Its stated goal was to “pave the way for truly beneficial and responsible AI.”
A few months later, a controversial contract between Google and the Pentagon was disclosed, causing an internal uproar in which employees accused Google of getting into “the business of war.”
Google’s Pentagon contract, known as Project Maven, “set alarm bells ringing” inside DeepMind, a former employee said. Afterward, Google published a set of principles to govern its work in AI, guidelines that were similar to the ethical charter that DeepMind had already set out internally, rankling some of DeepMind’s senior leadership, two former employees said.
Who will control the AI?
In April, Hassabis told employees in an all-hands meeting that negotiations to separate from Google had ended. DeepMind would maintain its existing status inside Alphabet. DeepMind’s future work would be overseen by Google’s Advanced Technology Review Council, which includes two DeepMind executives, Google’s AI chief Jeff Dean, and the legal SVP Kent Walker.
But the group’s yearslong battle to achieve more independence raises questions about its future within Google.
Google’s commitment to AI research has also come under question, after the company forced out two of its most senior AI ethics researchers. That led to an industry backlash and sowed doubt over whether it could allow truly independent research.
Ali Alkhatib, a fellow at the Center for Applied Data Ethics, told Insider that more public accountability was “desperately needed” to regulate the pursuit of AI by large tech companies.
For Google, its investment in DeepMind may be starting to pay off. Late last year, DeepMind announced a breakthrough to help scientists better understand the behavior of microscopic proteins, which has the potential to revolutionize drug discovery.
As for DeepMind, Hassabis is holding on to the belief that AI technology should not be controlled by a single corporation. Speaking at Tortoise’s Responsible AI Forum in June, he proposed a “world institute” of AI. Such a body might sit under the jurisdiction of the United Nations, Hassabis theorized, and could be filled with top researchers in the field.
“It’s much stronger if you lead by example,” he told the audience, “and I hope DeepMind can be part of that role-modeling for the industry.”